Temporal Attention Convolutional Neural Networks Based on LSTM-Encoder for Time Series Forecasting

• Published in: 2023 International Conference on Networks, Communications and Intelligent Computing (NCIC) pp. 51 - 54
• Main Authors: Zhou, Yantong, Chen, Ziwen, Xie, Anqi
• Format: Conference Proceeding
• Language: English
• Published: IEEE 17.11.2023
• Subjects: Analytical models; Data models; LSTM-encoder; Planning; Predictive models; TACN; Time series analysis; time series prediction; Wind farms; Wind power generation
• DOI: 10.1109/NCIC61838.2023.00014

Traditional time series prediction algorithms often struggle to model and forecast high-dimensional, complex data from real-world tasks. Here we present a temporal attention convolutional neural network (TACN) based on an LSTM encoder, which leverages attention mechanisms to process long sequences. For wind power generation scenarios, we combined the LSTM encoding structure with features engineered from sinusoidal and cosine laws of wind speed variation. Experimental comparisons with basic neural network benchmarks demonstrate the superior predictive performance of our TACN model for wind power forecasts. These results indicate potential for applying the approach to practical wind farm management and planning.